Visual analytics using artificial intelligence (multi-modality driver action recognition)
A report published by the National Highway Traffic Safety Administration (NHTSA) in the United States showed that up to 3522 people were killed due to distracted driving. Various driver monitoring system were developed to tackle this issue and potentially saving lives and increasing road safety,...
Saved in:
| 主要作者: | |
|---|---|
| 其他作者: | |
| 格式: | Final Year Project |
| 語言: | English |
| 出版: |
Nanyang Technological University
2024
|
| 主題: | |
| 在線閱讀: | https://hdl.handle.net/10356/176634 |
| 標簽: |
添加標簽
沒有標簽, 成為第一個標記此記錄!
|
| 總結: | A report published by the National Highway Traffic Safety Administration (NHTSA) in the
United States showed that up to 3522 people were killed due to distracted driving. Various
driver monitoring system were developed to tackle this issue and potentially saving lives and
increasing road safety, one such system includes a driver video action recognition system.
The project aims to develop a robust and stable driver action recognition model utilizing multimodality data streams, including RGB, IR and depth. A literature review was carried out to
determine suitable model and dataset for this project. Following model and dataset selection,
hyperparameters tuning is conducted to optimize VideoMAE V2 for improved accuracy and
efficiency on the Drive&Act (DAA) dataset. Various fusion learning technique were explored
and implemented into the system for evaluation. Early fusion achieves an average Top-1
accuracy of 82.40%, while late fusion obtains an average Top-1 accuracy of 84.30% on the test
set.
Overall, the project demonstrated the capability of incorporating early and late fusion methods
with VideoMAE V2 model to achieve satisfactory results. This suggests the potential
applicability of this model to different multi-modality action recognition tasks. Future work
explores alternative fusion techniques and expanding the model to other driver datasets. |
|---|